Machine Learning In Intelligent Fault Diagnosis

Learning is the most essential property for any intelligent system while machine learning is becoming one of the most profound research fields in AI. As an important applied branch in AI, to effectively acquire, transfer, deal with, regenerate and utilize diagnostic information has been the core of intelligent fault diagnosis (IFD) that exhibits an ability to precisely identify current status and predict future status of the system to be monitored. However, for the bottleneck of machine learning, it may be difficult to get an exact identifying result for some systems with strong uncertainty and randomicity or those with redundant or imperfect diagnostic information. Focusing on the key problem in IFD, machine learning, some important and unsolved topics are lucubrated systematically in this dissertation including the multi-representation of diagnostic symptoms knowledge, the compression of redundancy information and knowledge mining, fault pattern classification and trend prediction for the cases only very limited diagnostic information can be utilized. From the point of view of epistemology and generalized information theory, a new explanation of diagnosis based on information entropy that illustrates the flowing and transferring of Information and Knowledge is deducted mathematically. Four conclusions on diagnostic knowledge are summarized. The relationship among diagnostic signal, domain knowledge and machine learning is explained. Meanwhile, the relationship curve among diagnosis, cognitive ability and the amount of information for diagnosis is presented. Based on these results, the idea of Multi-Symptom Domains Comprehensive Feature Knowledge to represent diagnostic knowledge from multi-sides is put forwarded, which can overcome the imperfections caused by insufficient knowledge and lower cognitive ability for complex knowledge system.This paper presents and implements an effective data analysis approach with two functions of data compression that it can not only reduce the dimension of data by getting rid of the correlation among them but also remove the duplicated or proximately similar data. At the first step of the algorithm, the principal component analysis (PCA) approach is employed to reduce the dimension of data. Next, a modified immune clustering method inspired by the clonal section operation and immune network hypothesis of vertebrate's immune system is used to remove the unrepresentative samples according to the methodology of clustering but different to standard clustering approaches.The redefinition of affinity based on similarity measurement of principle component core is a creative modification to original algorithm. Plus other mending steps such as the normalization of antigen data and deleting duplicated data directly, a serials improvements make the new algorithm more efficient and applicable than the old one. Simulation experiments on the data from"Tennessee Eastman"plant that is popularly used in process control field proved the effectiveness of this algorithm. This dissertation emphasized the significance of generalization of learning machine for the performance of diagnosis when it is unavailable to acquire sufficient fault data. Normally, It is an ill-posed problem for traditional statistic learning methods that exhibitrules of infinity to solve diagnosis of small sample size data. Support vector machines (SVM) is a new general machine learning tools based on structural risk minimization principle (SRMP), which pursue the minimization of both empirical risk and confidence interval. Even if the size of learning sample is limited, SVM can perform interested generalization. Diagnosis based on SVM is presented and the basic implementing algorithm of this diagnostic approach is designed. Owing to that SVM is originally designed for binary classification, while most of diagnosis problems are multi-class cases, a new multi-class classification based on SVM named 2PTMC different to existing methods is presented. Thi...